In a study of 372 patients, the model closely matched results from a 70-gene genomic test and identified differences in long-term survival.

Agendia presented clinical validation data for an investigational artificial intelligence (AI) digital pathology model that infers MammaPrint 70-gene recurrence risk directly from routine hematoxylin and eosin (H&E) whole-slide images at the European Society for Medical Oncology (ESMO) Congress 2026 in Madrid.

The podium presentation, delivered by Agendia co-founder and chief research officer Laura van ’t Veer, PhD, evaluated whether deep learning algorithms applied to digitized standard biopsy slides could replicate the prognostic classification of the established tissue-based genomic assay in early-stage breast cancer.

“AI has been part of Agendia’s scientific foundation from the beginning and this study represents its continued evolution,” says Laura van ’t Veer, PhD, co-founder and chief research officer at Agendia, in a release. “For the first time, we are presenting our work exploring whether the biological information represented by MammaPrint can be recognized through AI applied to routinely generated pathology images.”

Model Architecture and Validation Cohort

The investigational pipeline was trained on 70,000 de-identified H&E whole-slide images using MammaPrint-derived supervision. The architecture incorporates tissue detection, classification, and feature extraction powered by self-supervised and supervised foundation models. The pipeline outputs a continuous recurrence score that is dichotomized into digital high-risk (D-High) or digital low-risk (D-Low) classifications.

Investigators validated the model using de-identified data from 372 early-stage breast cancer patients enrolled in the prospective, real-world RASTER study, which has a median follow-up of 10.3 years.

Across the validation cohort, the digital assay achieved a correlation coefficient of r = 0.954 with MammaPrint results, classifying 49% of patients as D-High and 51% as D-Low.

Survival Outcomes and Multivariable Performance

Long-term survival outcomes aligned with the digital risk classifications:

  • Patients classified as D-Low demonstrated a 10-year breast cancer-specific survival rate of 98.1%, compared to 86.7% for those classified as D-High (HR 0.12, p < 0.0001).
  • Among the subgroup of 274 patients with estrogen receptor-positive (ER+), HER2-negative tumors, 10-year breast cancer-specific survival was 98.1% for D-Low versus 84.7% for D-High (HR 0.10).
  • In a multivariable analysis, digital risk classification emerged as the only independent predictor of breast cancer-specific survival (HR 0.084, p = 0.0002), outperforming standard clinical factors. Similar predictive performance was documented for distant recurrence-free intervals.

“AI-driven digital pathology approaches have the potential to expand access to genomic information for breast cancer patients globally, particularly in scenarios where access to tissue-based genomics is limited,” says William Audeh, MD, MS, chief medical officer at Agendia, in a release. “This research brings together decades of genomic science with emerging digital pathology capabilities and reflects our commitment to developing innovative approaches to understanding tumor biology.”

At the congress, Agendia also presented two additional posters featuring data from its ongoing prospective, real-world FLEX Study.

ID 464502804 © goir | Dreamstime.com